Machine Learning-Based Traffic Prediction and Traffic-Aware Vehicle Routing Optimization

Abstract

Fixed-speed planning approaches widely used in urban logistics operations lead to unreliable delivery commitments by disregarding the dynamic traffic conditions of large cities. This study aims to develop an integrated pipeline that connects traffic speed prediction directly to road network cost calculation and vehicle routing optimization. A CatBoost gradient boosting algorithm was trained using approximately 19 million rows of 2024 traffic data obtained from the Istanbul Metropolitan Municipality Open Data Portal. The developed M2 model achieved R²=0.856 in temporal testing, while in the spatial holdout test — where the baseline model collapsed with R²=−0.340 — it maintained functionality with R²=0.418, demonstrating that geographic features contributed +0.758 to spatial generalization capacity. Consistent performance in the range of R²=0.863–0.884 was achieved in independent external validation conducted with data from January 2025. Model predictions were transferred to the Istanbul road network of 506,667 edges via Dijkstra's algorithm and provided as input to the Vehicle Routing Problem with Time Windows solved using OR-Tools. The dynamic validation analysis revealed a critical finding: while the static model predicted reaching 28 out of 30 points on time, simulation of this plan with traffic predictions showed that only 19 points could be reached. This 32% error rate concretely demonstrates that traffic-aware planning is indispensable for operational reliability.